Teamday
← Back

Compound AI: Why It Rebuilt Office for Finance Agents

AI Engineer
agentsenterprisebusinessknowledge-workstartup

How Compound AI Built Long-Running Agents for Buy-Side Finance

Yijia Liang spent about three years as a vice president specializing in late-stage private software investments, and "way too much time in Excel, PowerPoint, and late nights", before leaving to build tools to replace that job. The company, Compound AI (getcompound.ai), builds agents for private equity and hedge fund deal teams. The AI Engineer talk explains why the things that make personal agents work (reliable background tasks, proactivity, integrations with email and calendar) don't carry over to institutional finance. The finance version of a background task is turning a data room into a full set of investment committee models, slides and memos, a process that normally takes four weeks and several review rounds. Compound wants to compress it to a few days.

Multiplayer through the inbox: Deployments of Claude or ChatGPT in finance firms are single player today, Liang argues. Each person works in a plugin or web app, then downloads the output and emails it around for feedback. Compound puts the whole deal team on one thread with one agent that holds everyone's context. Every agent gets its own email inbox, because "for institutional finance, most workflows run through Outlook." Making that work is not trivial. Everyone on an NDA-bound thread needs authenticated read and write access to the agent, and the agent deliberately adds latency so that several quick replies don't trigger a burst of acknowledgments.

Rebuilding Office from scratch: Compound evaluated the Excel, PowerPoint and Word desktop plugins, computer use on virtual machines, open-source libraries like openpyxl and commercial engines. None could run many agent tasks in parallel at the quality firms need. So the team pointed coding agents at an objective function, recreate Office in full fidelity, and let them loop: "we ran agents in many loops, spend 1.5 trillion input output tokens and over 8 million lines of code to rewrite Office from scratch." For Excel they started from the open-source engine IronCalc, used Office as the oracle and benchmarked against real-world and internal files. The Excel replica alone is about 800,000 lines of Rust, and Rust renderers check that output is pixel-identical when downloaded. One-to-one parity matters because firms only trust what they can download and send.

Why a headless engine wins: Desktop Office is bound to one machine's CPU. A headless engine runs tasks in parallel and recalculates 20 MB workbooks about four times faster. Owning the editor also lets the agent check its own work. With a code library there is no renderer, so the model guesses at formatting. With Microsoft's interface you are limited to what Microsoft exposes. Compound owns the render, check and fix loop. Liang built the talk's slide deck by dropping comments on an agent's first draft, and says a team of five can work through a 100-page deck in an afternoon instead of weeks of revisions.

Swarms instead of vector databases: Deal rooms can hold 15,000 files, and single models reach 30 MB with two million formulas. For breadth, Compound borrows from coding agents. A lead agent slices the file directory, swarms of agents read files with grep-style tools and report back, and the loop repeats until the output is good enough. Every finding is cited to the source document. For depth, the engine turns a large workbook into a compact representation the model can read, instead of having the model write openpyxl code and print every cell.

5 Takeaways from Compound AI on Finance Agents

  • Finance needs multiplayer agents - Deal work involves a whole team, so the agent should sit in the shared thread, not in one analyst's plugin.
  • Email is the interface - Giving each agent an inbox meets Outlook-centric firms where they work, but it needs per-thread authentication and human-like pacing.
  • Coding agents make rebuilding infrastructure affordable - Compound used agent loops against Office as a judge to build its own Excel, PowerPoint and Word engines.
  • Own the renderer to own quality - A render, check and fix loop beats guessing at formatting through a code library.
  • Grep and swarms over indexing - For large deal rooms, coordinated multi-agent reading with citations replaced vector search for most uses.

From Copilot to Autopilot in Private Equity and Hedge Funds

Liang sees off-the-shelf tools already handling short tasks like spreading comps, while a 50-page IC deck or a model built exactly to a firm's standard still needs a human hand on every step. He compares finance today to coding after Opus 4.5 launched, when developers with enough checks and tests began delegating whole tasks. Compound's bet is that multiplayer threads, owned document engines and cited findings are the checks finance needs before analysts hand over a data room and trust the output. For another firm running its operations on agents, see Arca's AI-native wealth management.